4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
27 #include <sys/zfs_context.h>
28 #include <sys/spa_impl.h>
30 #include <sys/dmu_tx.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/metaslab.h>
34 #include <sys/metaslab_impl.h>
35 #include <sys/uberblock_impl.h>
38 #include <sys/bpobj.h>
39 #include <sys/dsl_pool.h>
40 #include <sys/dsl_synctask.h>
41 #include <sys/dsl_dir.h>
43 #include <sys/zfeature.h>
44 #include <sys/vdev_indirect_births.h>
45 #include <sys/vdev_indirect_mapping.h>
47 #include <sys/trace_vdev.h>
50 * This file contains the necessary logic to remove vdevs from a
51 * storage pool. Currently, the only devices that can be removed
52 * are log, cache, and spare devices; and top level vdevs from a pool
53 * w/o raidz or mirrors. (Note that members of a mirror can be removed
54 * by the detach operation.)
56 * Log vdevs are removed by evacuating them and then turning the vdev
57 * into a hole vdev while holding spa config locks.
59 * Top level vdevs are removed and converted into an indirect vdev via
60 * a multi-step process:
62 * - Disable allocations from this device (spa_vdev_remove_top).
64 * - From a new thread (spa_vdev_remove_thread), copy data from
65 * the removing vdev to a different vdev. The copy happens in open
66 * context (spa_vdev_copy_impl) and issues a sync task
67 * (vdev_mapping_sync) so the sync thread can update the partial
68 * indirect mappings in core and on disk.
70 * - If a free happens during a removal, it is freed from the
71 * removing vdev, and if it has already been copied, from the new
72 * location as well (free_from_removing_vdev).
74 * - After the removal is completed, the copy thread converts the vdev
75 * into an indirect vdev (vdev_remove_complete) before instructing
76 * the sync thread to destroy the space maps and finish the removal
77 * (spa_finish_removal).
80 typedef struct vdev_copy_arg
{
82 uint64_t vca_outstanding_bytes
;
88 * The maximum amount of memory we can use for outstanding i/o while
89 * doing a device removal. This determines how much i/o we can have
90 * in flight concurrently.
92 int zfs_remove_max_copy_bytes
= 64 * 1024 * 1024;
95 * The largest contiguous segment that we will attempt to allocate when
96 * removing a device. This can be no larger than SPA_MAXBLOCKSIZE. If
97 * there is a performance problem with attempting to allocate large blocks,
98 * consider decreasing this.
100 int zfs_remove_max_segment
= SPA_MAXBLOCKSIZE
;
102 #define VDEV_REMOVAL_ZAP_OBJS "lzap"
104 static void spa_vdev_remove_thread(void *arg
);
107 spa_sync_removing_state(spa_t
*spa
, dmu_tx_t
*tx
)
109 VERIFY0(zap_update(spa
->spa_dsl_pool
->dp_meta_objset
,
110 DMU_POOL_DIRECTORY_OBJECT
,
111 DMU_POOL_REMOVING
, sizeof (uint64_t),
112 sizeof (spa
->spa_removing_phys
) / sizeof (uint64_t),
113 &spa
->spa_removing_phys
, tx
));
117 spa_nvlist_lookup_by_guid(nvlist_t
**nvpp
, int count
, uint64_t target_guid
)
119 for (int i
= 0; i
< count
; i
++) {
121 fnvlist_lookup_uint64(nvpp
[i
], ZPOOL_CONFIG_GUID
);
123 if (guid
== target_guid
)
131 spa_vdev_remove_aux(nvlist_t
*config
, char *name
, nvlist_t
**dev
, int count
,
132 nvlist_t
*dev_to_remove
)
134 nvlist_t
**newdev
= NULL
;
137 newdev
= kmem_alloc((count
- 1) * sizeof (void *), KM_SLEEP
);
139 for (int i
= 0, j
= 0; i
< count
; i
++) {
140 if (dev
[i
] == dev_to_remove
)
142 VERIFY(nvlist_dup(dev
[i
], &newdev
[j
++], KM_SLEEP
) == 0);
145 VERIFY(nvlist_remove(config
, name
, DATA_TYPE_NVLIST_ARRAY
) == 0);
146 VERIFY(nvlist_add_nvlist_array(config
, name
, newdev
, count
- 1) == 0);
148 for (int i
= 0; i
< count
- 1; i
++)
149 nvlist_free(newdev
[i
]);
152 kmem_free(newdev
, (count
- 1) * sizeof (void *));
155 static spa_vdev_removal_t
*
156 spa_vdev_removal_create(vdev_t
*vd
)
158 spa_vdev_removal_t
*svr
= kmem_zalloc(sizeof (*svr
), KM_SLEEP
);
159 mutex_init(&svr
->svr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
160 cv_init(&svr
->svr_cv
, NULL
, CV_DEFAULT
, NULL
);
161 svr
->svr_allocd_segs
= range_tree_create(NULL
, NULL
);
162 svr
->svr_vdev_id
= vd
->vdev_id
;
164 for (int i
= 0; i
< TXG_SIZE
; i
++) {
165 svr
->svr_frees
[i
] = range_tree_create(NULL
, NULL
);
166 list_create(&svr
->svr_new_segments
[i
],
167 sizeof (vdev_indirect_mapping_entry_t
),
168 offsetof(vdev_indirect_mapping_entry_t
, vime_node
));
175 spa_vdev_removal_destroy(spa_vdev_removal_t
*svr
)
177 for (int i
= 0; i
< TXG_SIZE
; i
++) {
178 ASSERT0(svr
->svr_bytes_done
[i
]);
179 ASSERT0(svr
->svr_max_offset_to_sync
[i
]);
180 range_tree_destroy(svr
->svr_frees
[i
]);
181 list_destroy(&svr
->svr_new_segments
[i
]);
184 range_tree_destroy(svr
->svr_allocd_segs
);
185 mutex_destroy(&svr
->svr_lock
);
186 cv_destroy(&svr
->svr_cv
);
187 kmem_free(svr
, sizeof (*svr
));
191 * This is called as a synctask in the txg in which we will mark this vdev
192 * as removing (in the config stored in the MOS).
194 * It begins the evacuation of a toplevel vdev by:
195 * - initializing the spa_removing_phys which tracks this removal
196 * - computing the amount of space to remove for accounting purposes
197 * - dirtying all dbufs in the spa_config_object
198 * - creating the spa_vdev_removal
199 * - starting the spa_vdev_remove_thread
202 vdev_remove_initiate_sync(void *arg
, dmu_tx_t
*tx
)
204 int vdev_id
= (uintptr_t)arg
;
205 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
206 vdev_t
*vd
= vdev_lookup_top(spa
, vdev_id
);
207 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
208 objset_t
*mos
= spa
->spa_dsl_pool
->dp_meta_objset
;
209 spa_vdev_removal_t
*svr
= NULL
;
210 ASSERTV(uint64_t txg
= dmu_tx_get_txg(tx
));
212 ASSERT3P(vd
->vdev_ops
, !=, &vdev_raidz_ops
);
213 svr
= spa_vdev_removal_create(vd
);
215 ASSERT(vd
->vdev_removing
);
216 ASSERT3P(vd
->vdev_indirect_mapping
, ==, NULL
);
218 spa_feature_incr(spa
, SPA_FEATURE_DEVICE_REMOVAL
, tx
);
219 if (spa_feature_is_enabled(spa
, SPA_FEATURE_OBSOLETE_COUNTS
)) {
221 * By activating the OBSOLETE_COUNTS feature, we prevent
222 * the pool from being downgraded and ensure that the
223 * refcounts are precise.
225 spa_feature_incr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
227 VERIFY0(zap_add(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
228 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE
, sizeof (one
), 1,
230 ASSERT3U(vdev_obsolete_counts_are_precise(vd
), !=, 0);
233 vic
->vic_mapping_object
= vdev_indirect_mapping_alloc(mos
, tx
);
234 vd
->vdev_indirect_mapping
=
235 vdev_indirect_mapping_open(mos
, vic
->vic_mapping_object
);
236 vic
->vic_births_object
= vdev_indirect_births_alloc(mos
, tx
);
237 vd
->vdev_indirect_births
=
238 vdev_indirect_births_open(mos
, vic
->vic_births_object
);
239 spa
->spa_removing_phys
.sr_removing_vdev
= vd
->vdev_id
;
240 spa
->spa_removing_phys
.sr_start_time
= gethrestime_sec();
241 spa
->spa_removing_phys
.sr_end_time
= 0;
242 spa
->spa_removing_phys
.sr_state
= DSS_SCANNING
;
243 spa
->spa_removing_phys
.sr_to_copy
= 0;
244 spa
->spa_removing_phys
.sr_copied
= 0;
247 * Note: We can't use vdev_stat's vs_alloc for sr_to_copy, because
248 * there may be space in the defer tree, which is free, but still
249 * counted in vs_alloc.
251 for (uint64_t i
= 0; i
< vd
->vdev_ms_count
; i
++) {
252 metaslab_t
*ms
= vd
->vdev_ms
[i
];
253 if (ms
->ms_sm
== NULL
)
257 * Sync tasks happen before metaslab_sync(), therefore
258 * smp_alloc and sm_alloc must be the same.
260 ASSERT3U(space_map_allocated(ms
->ms_sm
), ==,
261 ms
->ms_sm
->sm_phys
->smp_alloc
);
263 spa
->spa_removing_phys
.sr_to_copy
+=
264 space_map_allocated(ms
->ms_sm
);
267 * Space which we are freeing this txg does not need to
270 spa
->spa_removing_phys
.sr_to_copy
-=
271 range_tree_space(ms
->ms_freeingtree
);
273 ASSERT0(range_tree_space(ms
->ms_freedtree
));
274 for (int t
= 0; t
< TXG_SIZE
; t
++)
275 ASSERT0(range_tree_space(ms
->ms_alloctree
[t
]));
279 * Sync tasks are called before metaslab_sync(), so there should
280 * be no already-synced metaslabs in the TXG_CLEAN list.
282 ASSERT3P(txg_list_head(&vd
->vdev_ms_list
, TXG_CLEAN(txg
)), ==, NULL
);
284 spa_sync_removing_state(spa
, tx
);
287 * All blocks that we need to read the most recent mapping must be
288 * stored on concrete vdevs. Therefore, we must dirty anything that
289 * is read before spa_remove_init(). Specifically, the
290 * spa_config_object. (Note that although we already modified the
291 * spa_config_object in spa_sync_removing_state, that may not have
292 * modified all blocks of the object.)
294 dmu_object_info_t doi
;
295 VERIFY0(dmu_object_info(mos
, DMU_POOL_DIRECTORY_OBJECT
, &doi
));
296 for (uint64_t offset
= 0; offset
< doi
.doi_max_offset
; ) {
298 VERIFY0(dmu_buf_hold(mos
, DMU_POOL_DIRECTORY_OBJECT
,
299 offset
, FTAG
, &dbuf
, 0));
300 dmu_buf_will_dirty(dbuf
, tx
);
301 offset
+= dbuf
->db_size
;
302 dmu_buf_rele(dbuf
, FTAG
);
306 * Now that we've allocated the im_object, dirty the vdev to ensure
307 * that the object gets written to the config on disk.
309 vdev_config_dirty(vd
);
311 zfs_dbgmsg("starting removal thread for vdev %llu (%p) in txg %llu "
312 "im_obj=%llu", vd
->vdev_id
, vd
, dmu_tx_get_txg(tx
),
313 vic
->vic_mapping_object
);
315 spa_history_log_internal(spa
, "vdev remove started", tx
,
316 "%s vdev %llu %s", spa_name(spa
), vd
->vdev_id
,
317 (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
319 * Setting spa_vdev_removal causes subsequent frees to call
320 * free_from_removing_vdev(). Note that we don't need any locking
321 * because we are the sync thread, and metaslab_free_impl() is only
322 * called from syncing context (potentially from a zio taskq thread,
323 * but in any case only when there are outstanding free i/os, which
326 ASSERT3P(spa
->spa_vdev_removal
, ==, NULL
);
327 spa
->spa_vdev_removal
= svr
;
328 svr
->svr_thread
= thread_create(NULL
, 0,
329 spa_vdev_remove_thread
, spa
, 0, &p0
, TS_RUN
, minclsyspri
);
333 * When we are opening a pool, we must read the mapping for each
334 * indirect vdev in order from most recently removed to least
335 * recently removed. We do this because the blocks for the mapping
336 * of older indirect vdevs may be stored on more recently removed vdevs.
337 * In order to read each indirect mapping object, we must have
338 * initialized all more recently removed vdevs.
341 spa_remove_init(spa_t
*spa
)
345 error
= zap_lookup(spa
->spa_dsl_pool
->dp_meta_objset
,
346 DMU_POOL_DIRECTORY_OBJECT
,
347 DMU_POOL_REMOVING
, sizeof (uint64_t),
348 sizeof (spa
->spa_removing_phys
) / sizeof (uint64_t),
349 &spa
->spa_removing_phys
);
351 if (error
== ENOENT
) {
352 spa
->spa_removing_phys
.sr_state
= DSS_NONE
;
353 spa
->spa_removing_phys
.sr_removing_vdev
= -1;
354 spa
->spa_removing_phys
.sr_prev_indirect_vdev
= -1;
356 } else if (error
!= 0) {
360 if (spa
->spa_removing_phys
.sr_state
== DSS_SCANNING
) {
362 * We are currently removing a vdev. Create and
363 * initialize a spa_vdev_removal_t from the bonus
364 * buffer of the removing vdevs vdev_im_object, and
365 * initialize its partial mapping.
367 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
368 vdev_t
*vd
= vdev_lookup_top(spa
,
369 spa
->spa_removing_phys
.sr_removing_vdev
);
372 spa_config_exit(spa
, SCL_STATE
, FTAG
);
376 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
378 ASSERT(vdev_is_concrete(vd
));
379 spa_vdev_removal_t
*svr
= spa_vdev_removal_create(vd
);
380 ASSERT3U(svr
->svr_vdev_id
, ==, vd
->vdev_id
);
381 ASSERT(vd
->vdev_removing
);
383 vd
->vdev_indirect_mapping
= vdev_indirect_mapping_open(
384 spa
->spa_meta_objset
, vic
->vic_mapping_object
);
385 vd
->vdev_indirect_births
= vdev_indirect_births_open(
386 spa
->spa_meta_objset
, vic
->vic_births_object
);
387 spa_config_exit(spa
, SCL_STATE
, FTAG
);
389 spa
->spa_vdev_removal
= svr
;
392 spa_config_enter(spa
, SCL_STATE
, FTAG
, RW_READER
);
393 uint64_t indirect_vdev_id
=
394 spa
->spa_removing_phys
.sr_prev_indirect_vdev
;
395 while (indirect_vdev_id
!= UINT64_MAX
) {
396 vdev_t
*vd
= vdev_lookup_top(spa
, indirect_vdev_id
);
397 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
399 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
400 vd
->vdev_indirect_mapping
= vdev_indirect_mapping_open(
401 spa
->spa_meta_objset
, vic
->vic_mapping_object
);
402 vd
->vdev_indirect_births
= vdev_indirect_births_open(
403 spa
->spa_meta_objset
, vic
->vic_births_object
);
405 indirect_vdev_id
= vic
->vic_prev_indirect_vdev
;
407 spa_config_exit(spa
, SCL_STATE
, FTAG
);
410 * Now that we've loaded all the indirect mappings, we can allow
411 * reads from other blocks (e.g. via predictive prefetch).
413 spa
->spa_indirect_vdevs_loaded
= B_TRUE
;
418 spa_restart_removal(spa_t
*spa
)
420 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
426 * In general when this function is called there is no
427 * removal thread running. The only scenario where this
428 * is not true is during spa_import() where this function
429 * is called twice [once from spa_import_impl() and
430 * spa_async_resume()]. Thus, in the scenario where we
431 * import a pool that has an ongoing removal we don't
432 * want to spawn a second thread.
434 if (svr
->svr_thread
!= NULL
)
437 if (!spa_writeable(spa
))
440 zfs_dbgmsg("restarting removal of %llu", svr
->svr_vdev_id
);
441 svr
->svr_thread
= thread_create(NULL
, 0, spa_vdev_remove_thread
, spa
,
442 0, &p0
, TS_RUN
, minclsyspri
);
446 * Process freeing from a device which is in the middle of being removed.
447 * We must handle this carefully so that we attempt to copy freed data,
448 * and we correctly free already-copied data.
451 free_from_removing_vdev(vdev_t
*vd
, uint64_t offset
, uint64_t size
,
454 spa_t
*spa
= vd
->vdev_spa
;
455 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
456 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
457 uint64_t max_offset_yet
= 0;
459 ASSERT(vd
->vdev_indirect_config
.vic_mapping_object
!= 0);
460 ASSERT3U(vd
->vdev_indirect_config
.vic_mapping_object
, ==,
461 vdev_indirect_mapping_object(vim
));
462 ASSERT3U(vd
->vdev_id
, ==, svr
->svr_vdev_id
);
463 ASSERT3U(spa_syncing_txg(spa
), ==, txg
);
465 mutex_enter(&svr
->svr_lock
);
468 * Remove the segment from the removing vdev's spacemap. This
469 * ensures that we will not attempt to copy this space (if the
470 * removal thread has not yet visited it), and also ensures
471 * that we know what is actually allocated on the new vdevs
472 * (needed if we cancel the removal).
474 * Note: we must do the metaslab_free_concrete() with the svr_lock
475 * held, so that the remove_thread can not load this metaslab and then
476 * visit this offset between the time that we metaslab_free_concrete()
477 * and when we check to see if it has been visited.
479 metaslab_free_concrete(vd
, offset
, size
, txg
);
481 uint64_t synced_size
= 0;
482 uint64_t synced_offset
= 0;
483 uint64_t max_offset_synced
= vdev_indirect_mapping_max_offset(vim
);
484 if (offset
< max_offset_synced
) {
486 * The mapping for this offset is already on disk.
487 * Free from the new location.
489 * Note that we use svr_max_synced_offset because it is
490 * updated atomically with respect to the in-core mapping.
491 * By contrast, vim_max_offset is not.
493 * This block may be split between a synced entry and an
494 * in-flight or unvisited entry. Only process the synced
495 * portion of it here.
497 synced_size
= MIN(size
, max_offset_synced
- offset
);
498 synced_offset
= offset
;
500 ASSERT3U(max_offset_yet
, <=, max_offset_synced
);
501 max_offset_yet
= max_offset_synced
;
503 DTRACE_PROBE3(remove__free__synced
,
506 uint64_t, synced_size
);
509 offset
+= synced_size
;
513 * Look at all in-flight txgs starting from the currently syncing one
514 * and see if a section of this free is being copied. By starting from
515 * this txg and iterating forward, we might find that this region
516 * was copied in two different txgs and handle it appropriately.
518 for (int i
= 0; i
< TXG_CONCURRENT_STATES
; i
++) {
519 int txgoff
= (txg
+ i
) & TXG_MASK
;
520 if (size
> 0 && offset
< svr
->svr_max_offset_to_sync
[txgoff
]) {
522 * The mapping for this offset is in flight, and
523 * will be synced in txg+i.
525 uint64_t inflight_size
= MIN(size
,
526 svr
->svr_max_offset_to_sync
[txgoff
] - offset
);
528 DTRACE_PROBE4(remove__free__inflight
,
531 uint64_t, inflight_size
,
535 * We copy data in order of increasing offset.
536 * Therefore the max_offset_to_sync[] must increase
537 * (or be zero, indicating that nothing is being
538 * copied in that txg).
540 if (svr
->svr_max_offset_to_sync
[txgoff
] != 0) {
541 ASSERT3U(svr
->svr_max_offset_to_sync
[txgoff
],
544 svr
->svr_max_offset_to_sync
[txgoff
];
548 * We've already committed to copying this segment:
549 * we have allocated space elsewhere in the pool for
550 * it and have an IO outstanding to copy the data. We
551 * cannot free the space before the copy has
552 * completed, or else the copy IO might overwrite any
553 * new data. To free that space, we record the
554 * segment in the appropriate svr_frees tree and free
555 * the mapped space later, in the txg where we have
556 * completed the copy and synced the mapping (see
557 * vdev_mapping_sync).
559 range_tree_add(svr
->svr_frees
[txgoff
],
560 offset
, inflight_size
);
561 size
-= inflight_size
;
562 offset
+= inflight_size
;
565 * This space is already accounted for as being
566 * done, because it is being copied in txg+i.
567 * However, if i!=0, then it is being copied in
568 * a future txg. If we crash after this txg
569 * syncs but before txg+i syncs, then the space
570 * will be free. Therefore we must account
571 * for the space being done in *this* txg
572 * (when it is freed) rather than the future txg
573 * (when it will be copied).
575 ASSERT3U(svr
->svr_bytes_done
[txgoff
], >=,
577 svr
->svr_bytes_done
[txgoff
] -= inflight_size
;
578 svr
->svr_bytes_done
[txg
& TXG_MASK
] += inflight_size
;
581 ASSERT0(svr
->svr_max_offset_to_sync
[TXG_CLEAN(txg
) & TXG_MASK
]);
585 * The copy thread has not yet visited this offset. Ensure
589 DTRACE_PROBE3(remove__free__unvisited
,
594 if (svr
->svr_allocd_segs
!= NULL
)
595 range_tree_clear(svr
->svr_allocd_segs
, offset
, size
);
598 * Since we now do not need to copy this data, for
599 * accounting purposes we have done our job and can count
602 svr
->svr_bytes_done
[txg
& TXG_MASK
] += size
;
604 mutex_exit(&svr
->svr_lock
);
607 * Now that we have dropped svr_lock, process the synced portion
610 if (synced_size
> 0) {
611 vdev_indirect_mark_obsolete(vd
, synced_offset
, synced_size
,
614 * Note: this can only be called from syncing context,
615 * and the vdev_indirect_mapping is only changed from the
616 * sync thread, so we don't need svr_lock while doing
617 * metaslab_free_impl_cb.
619 vdev_indirect_ops
.vdev_op_remap(vd
, synced_offset
, synced_size
,
620 metaslab_free_impl_cb
, &txg
);
625 * Stop an active removal and update the spa_removing phys.
628 spa_finish_removal(spa_t
*spa
, dsl_scan_state_t state
, dmu_tx_t
*tx
)
630 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
631 ASSERT3U(dmu_tx_get_txg(tx
), ==, spa_syncing_txg(spa
));
633 /* Ensure the removal thread has completed before we free the svr. */
634 spa_vdev_remove_suspend(spa
);
636 ASSERT(state
== DSS_FINISHED
|| state
== DSS_CANCELED
);
638 if (state
== DSS_FINISHED
) {
639 spa_removing_phys_t
*srp
= &spa
->spa_removing_phys
;
640 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
641 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
643 if (srp
->sr_prev_indirect_vdev
!= UINT64_MAX
) {
645 pvd
= vdev_lookup_top(spa
,
646 srp
->sr_prev_indirect_vdev
);
647 ASSERT3P(pvd
->vdev_ops
, ==, &vdev_indirect_ops
);
650 vic
->vic_prev_indirect_vdev
= srp
->sr_prev_indirect_vdev
;
651 srp
->sr_prev_indirect_vdev
= vd
->vdev_id
;
653 spa
->spa_removing_phys
.sr_state
= state
;
654 spa
->spa_removing_phys
.sr_end_time
= gethrestime_sec();
656 spa
->spa_vdev_removal
= NULL
;
657 spa_vdev_removal_destroy(svr
);
659 spa_sync_removing_state(spa
, tx
);
661 vdev_config_dirty(spa
->spa_root_vdev
);
665 free_mapped_segment_cb(void *arg
, uint64_t offset
, uint64_t size
)
668 vdev_indirect_mark_obsolete(vd
, offset
, size
,
669 vd
->vdev_spa
->spa_syncing_txg
);
670 vdev_indirect_ops
.vdev_op_remap(vd
, offset
, size
,
671 metaslab_free_impl_cb
, &vd
->vdev_spa
->spa_syncing_txg
);
675 * On behalf of the removal thread, syncs an incremental bit more of
676 * the indirect mapping to disk and updates the in-memory mapping.
677 * Called as a sync task in every txg that the removal thread makes progress.
680 vdev_mapping_sync(void *arg
, dmu_tx_t
*tx
)
682 spa_vdev_removal_t
*svr
= arg
;
683 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
684 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
685 ASSERTV(vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
);
686 uint64_t txg
= dmu_tx_get_txg(tx
);
687 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
689 ASSERT(vic
->vic_mapping_object
!= 0);
690 ASSERT3U(txg
, ==, spa_syncing_txg(spa
));
692 vdev_indirect_mapping_add_entries(vim
,
693 &svr
->svr_new_segments
[txg
& TXG_MASK
], tx
);
694 vdev_indirect_births_add_entry(vd
->vdev_indirect_births
,
695 vdev_indirect_mapping_max_offset(vim
), dmu_tx_get_txg(tx
), tx
);
698 * Free the copied data for anything that was freed while the
699 * mapping entries were in flight.
701 mutex_enter(&svr
->svr_lock
);
702 range_tree_vacate(svr
->svr_frees
[txg
& TXG_MASK
],
703 free_mapped_segment_cb
, vd
);
704 ASSERT3U(svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
], >=,
705 vdev_indirect_mapping_max_offset(vim
));
706 svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] = 0;
707 mutex_exit(&svr
->svr_lock
);
709 spa_sync_removing_state(spa
, tx
);
713 * All reads and writes associated with a call to spa_vdev_copy_segment()
717 spa_vdev_copy_nullzio_done(zio_t
*zio
)
719 spa_config_exit(zio
->io_spa
, SCL_STATE
, zio
->io_spa
);
723 * The write of the new location is done.
726 spa_vdev_copy_segment_write_done(zio_t
*zio
)
728 vdev_copy_arg_t
*vca
= zio
->io_private
;
730 abd_free(zio
->io_abd
);
732 mutex_enter(&vca
->vca_lock
);
733 vca
->vca_outstanding_bytes
-= zio
->io_size
;
734 cv_signal(&vca
->vca_cv
);
735 mutex_exit(&vca
->vca_lock
);
739 * The read of the old location is done. The parent zio is the write to
740 * the new location. Allow it to start.
743 spa_vdev_copy_segment_read_done(zio_t
*zio
)
745 zio_nowait(zio_unique_parent(zio
));
749 * If the old and new vdevs are mirrors, we will read both sides of the old
750 * mirror, and write each copy to the corresponding side of the new mirror.
751 * If the old and new vdevs have a different number of children, we will do
752 * this as best as possible. Since we aren't verifying checksums, this
753 * ensures that as long as there's a good copy of the data, we'll have a
754 * good copy after the removal, even if there's silent damage to one side
755 * of the mirror. If we're removing a mirror that has some silent damage,
756 * we'll have exactly the same damage in the new location (assuming that
757 * the new location is also a mirror).
759 * We accomplish this by creating a tree of zio_t's, with as many writes as
760 * there are "children" of the new vdev (a non-redundant vdev counts as one
761 * child, a 2-way mirror has 2 children, etc). Each write has an associated
762 * read from a child of the old vdev. Typically there will be the same
763 * number of children of the old and new vdevs. However, if there are more
764 * children of the new vdev, some child(ren) of the old vdev will be issued
765 * multiple reads. If there are more children of the old vdev, some copies
768 * For example, the tree of zio_t's for a 2-way mirror is:
772 * write(new vdev, child 0) write(new vdev, child 1)
774 * read(old vdev, child 0) read(old vdev, child 1)
776 * Child zio's complete before their parents complete. However, zio's
777 * created with zio_vdev_child_io() may be issued before their children
778 * complete. In this case we need to make sure that the children (reads)
779 * complete before the parents (writes) are *issued*. We do this by not
780 * calling zio_nowait() on each write until its corresponding read has
783 * The spa_config_lock must be held while zio's created by
784 * zio_vdev_child_io() are in progress, to ensure that the vdev tree does
785 * not change (e.g. due to a concurrent "zpool attach/detach"). The "null"
786 * zio is needed to release the spa_config_lock after all the reads and
787 * writes complete. (Note that we can't grab the config lock for each read,
788 * because it is not reentrant - we could deadlock with a thread waiting
792 spa_vdev_copy_one_child(vdev_copy_arg_t
*vca
, zio_t
*nzio
,
793 vdev_t
*source_vd
, uint64_t source_offset
,
794 vdev_t
*dest_child_vd
, uint64_t dest_offset
, int dest_id
, uint64_t size
)
796 ASSERT3U(spa_config_held(nzio
->io_spa
, SCL_ALL
, RW_READER
), !=, 0);
798 mutex_enter(&vca
->vca_lock
);
799 vca
->vca_outstanding_bytes
+= size
;
800 mutex_exit(&vca
->vca_lock
);
802 abd_t
*abd
= abd_alloc_for_io(size
, B_FALSE
);
804 vdev_t
*source_child_vd
;
805 if (source_vd
->vdev_ops
== &vdev_mirror_ops
&& dest_id
!= -1) {
807 * Source and dest are both mirrors. Copy from the same
808 * child id as we are copying to (wrapping around if there
809 * are more dest children than source children).
812 source_vd
->vdev_child
[dest_id
% source_vd
->vdev_children
];
814 source_child_vd
= source_vd
;
817 zio_t
*write_zio
= zio_vdev_child_io(nzio
, NULL
,
818 dest_child_vd
, dest_offset
, abd
, size
,
819 ZIO_TYPE_WRITE
, ZIO_PRIORITY_REMOVAL
,
821 spa_vdev_copy_segment_write_done
, vca
);
823 zio_nowait(zio_vdev_child_io(write_zio
, NULL
,
824 source_child_vd
, source_offset
, abd
, size
,
825 ZIO_TYPE_READ
, ZIO_PRIORITY_REMOVAL
,
827 spa_vdev_copy_segment_read_done
, vca
));
831 * Allocate a new location for this segment, and create the zio_t's to
832 * read from the old location and write to the new location.
835 spa_vdev_copy_segment(vdev_t
*vd
, uint64_t start
, uint64_t size
, uint64_t txg
,
836 vdev_copy_arg_t
*vca
, zio_alloc_list_t
*zal
)
838 metaslab_group_t
*mg
= vd
->vdev_mg
;
839 spa_t
*spa
= vd
->vdev_spa
;
840 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
841 vdev_indirect_mapping_entry_t
*entry
;
844 ASSERT3U(size
, <=, SPA_MAXBLOCKSIZE
);
846 int error
= metaslab_alloc_dva(spa
, mg
->mg_class
, size
,
847 &dst
, 0, NULL
, txg
, 0, zal
);
852 * We can't have any padding of the allocated size, otherwise we will
853 * misunderstand what's allocated, and the size of the mapping.
854 * The caller ensures this will be true by passing in a size that is
855 * aligned to the worst (highest) ashift in the pool.
857 ASSERT3U(DVA_GET_ASIZE(&dst
), ==, size
);
859 entry
= kmem_zalloc(sizeof (vdev_indirect_mapping_entry_t
), KM_SLEEP
);
860 DVA_MAPPING_SET_SRC_OFFSET(&entry
->vime_mapping
, start
);
861 entry
->vime_mapping
.vimep_dst
= dst
;
864 * See comment before spa_vdev_copy_one_child().
866 spa_config_enter(spa
, SCL_STATE
, spa
, RW_READER
);
867 zio_t
*nzio
= zio_null(spa
->spa_txg_zio
[txg
& TXG_MASK
], spa
, NULL
,
868 spa_vdev_copy_nullzio_done
, NULL
, 0);
869 vdev_t
*dest_vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dst
));
870 if (dest_vd
->vdev_ops
== &vdev_mirror_ops
) {
871 for (int i
= 0; i
< dest_vd
->vdev_children
; i
++) {
872 vdev_t
*child
= dest_vd
->vdev_child
[i
];
873 spa_vdev_copy_one_child(vca
, nzio
, vd
, start
,
874 child
, DVA_GET_OFFSET(&dst
), i
, size
);
877 spa_vdev_copy_one_child(vca
, nzio
, vd
, start
,
878 dest_vd
, DVA_GET_OFFSET(&dst
), -1, size
);
882 list_insert_tail(&svr
->svr_new_segments
[txg
& TXG_MASK
], entry
);
883 ASSERT3U(start
+ size
, <=, vd
->vdev_ms_count
<< vd
->vdev_ms_shift
);
884 vdev_dirty(vd
, 0, NULL
, txg
);
890 * Complete the removal of a toplevel vdev. This is called as a
891 * synctask in the same txg that we will sync out the new config (to the
892 * MOS object) which indicates that this vdev is indirect.
895 vdev_remove_complete_sync(void *arg
, dmu_tx_t
*tx
)
897 spa_vdev_removal_t
*svr
= arg
;
898 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
899 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
901 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
903 for (int i
= 0; i
< TXG_SIZE
; i
++) {
904 ASSERT0(svr
->svr_bytes_done
[i
]);
907 ASSERT3U(spa
->spa_removing_phys
.sr_copied
, ==,
908 spa
->spa_removing_phys
.sr_to_copy
);
910 vdev_destroy_spacemaps(vd
, tx
);
912 /* destroy leaf zaps, if any */
913 ASSERT3P(svr
->svr_zaplist
, !=, NULL
);
914 for (nvpair_t
*pair
= nvlist_next_nvpair(svr
->svr_zaplist
, NULL
);
916 pair
= nvlist_next_nvpair(svr
->svr_zaplist
, pair
)) {
917 vdev_destroy_unlink_zap(vd
, fnvpair_value_uint64(pair
), tx
);
919 fnvlist_free(svr
->svr_zaplist
);
921 spa_finish_removal(dmu_tx_pool(tx
)->dp_spa
, DSS_FINISHED
, tx
);
922 /* vd->vdev_path is not available here */
923 spa_history_log_internal(spa
, "vdev remove completed", tx
,
924 "%s vdev %llu", spa_name(spa
), vd
->vdev_id
);
928 vdev_remove_enlist_zaps(vdev_t
*vd
, nvlist_t
*zlist
)
930 ASSERT3P(zlist
, !=, NULL
);
931 ASSERT3P(vd
->vdev_ops
, !=, &vdev_raidz_ops
);
933 if (vd
->vdev_leaf_zap
!= 0) {
935 (void) snprintf(zkey
, sizeof (zkey
), "%s-%llu",
936 VDEV_REMOVAL_ZAP_OBJS
, (u_longlong_t
)vd
->vdev_leaf_zap
);
937 fnvlist_add_uint64(zlist
, zkey
, vd
->vdev_leaf_zap
);
940 for (uint64_t id
= 0; id
< vd
->vdev_children
; id
++) {
941 vdev_remove_enlist_zaps(vd
->vdev_child
[id
], zlist
);
946 vdev_remove_replace_with_indirect(vdev_t
*vd
, uint64_t txg
)
950 spa_t
*spa
= vd
->vdev_spa
;
951 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
954 * First, build a list of leaf zaps to be destroyed.
955 * This is passed to the sync context thread,
956 * which does the actual unlinking.
958 svr
->svr_zaplist
= fnvlist_alloc();
959 vdev_remove_enlist_zaps(vd
, svr
->svr_zaplist
);
961 ivd
= vdev_add_parent(vd
, &vdev_indirect_ops
);
962 ivd
->vdev_removing
= 0;
964 vd
->vdev_leaf_zap
= 0;
966 vdev_remove_child(ivd
, vd
);
967 vdev_compact_children(ivd
);
969 ASSERT(!list_link_active(&vd
->vdev_state_dirty_node
));
971 tx
= dmu_tx_create_assigned(spa
->spa_dsl_pool
, txg
);
972 dsl_sync_task_nowait(spa
->spa_dsl_pool
, vdev_remove_complete_sync
, svr
,
973 0, ZFS_SPACE_CHECK_NONE
, tx
);
977 * Indicate that this thread has exited.
978 * After this, we can not use svr.
980 mutex_enter(&svr
->svr_lock
);
981 svr
->svr_thread
= NULL
;
982 cv_broadcast(&svr
->svr_cv
);
983 mutex_exit(&svr
->svr_lock
);
987 * Complete the removal of a toplevel vdev. This is called in open
988 * context by the removal thread after we have copied all vdev's data.
991 vdev_remove_complete(spa_t
*spa
)
996 * Wait for any deferred frees to be synced before we call
997 * vdev_metaslab_fini()
999 txg_wait_synced(spa
->spa_dsl_pool
, 0);
1000 txg
= spa_vdev_enter(spa
);
1001 vdev_t
*vd
= vdev_lookup_top(spa
, spa
->spa_vdev_removal
->svr_vdev_id
);
1003 sysevent_t
*ev
= spa_event_create(spa
, vd
, NULL
,
1004 ESC_ZFS_VDEV_REMOVE_DEV
);
1006 zfs_dbgmsg("finishing device removal for vdev %llu in txg %llu",
1010 * Discard allocation state.
1012 if (vd
->vdev_mg
!= NULL
) {
1013 vdev_metaslab_fini(vd
);
1014 metaslab_group_destroy(vd
->vdev_mg
);
1017 ASSERT0(vd
->vdev_stat
.vs_space
);
1018 ASSERT0(vd
->vdev_stat
.vs_dspace
);
1020 vdev_remove_replace_with_indirect(vd
, txg
);
1023 * We now release the locks, allowing spa_sync to run and finish the
1024 * removal via vdev_remove_complete_sync in syncing context.
1026 * Note that we hold on to the vdev_t that has been replaced. Since
1027 * it isn't part of the vdev tree any longer, it can't be concurrently
1028 * manipulated, even while we don't have the config lock.
1030 (void) spa_vdev_exit(spa
, NULL
, txg
, 0);
1033 * Top ZAP should have been transferred to the indirect vdev in
1034 * vdev_remove_replace_with_indirect.
1036 ASSERT0(vd
->vdev_top_zap
);
1039 * Leaf ZAP should have been moved in vdev_remove_replace_with_indirect.
1041 ASSERT0(vd
->vdev_leaf_zap
);
1043 txg
= spa_vdev_enter(spa
);
1044 (void) vdev_label_init(vd
, 0, VDEV_LABEL_REMOVE
);
1046 * Request to update the config and the config cachefile.
1048 vdev_config_dirty(spa
->spa_root_vdev
);
1049 (void) spa_vdev_exit(spa
, vd
, txg
, 0);
1056 * Evacuates a segment of size at most max_alloc from the vdev
1057 * via repeated calls to spa_vdev_copy_segment. If an allocation
1058 * fails, the pool is probably too fragmented to handle such a
1059 * large size, so decrease max_alloc so that the caller will not try
1060 * this size again this txg.
1063 spa_vdev_copy_impl(vdev_t
*vd
, spa_vdev_removal_t
*svr
, vdev_copy_arg_t
*vca
,
1064 uint64_t *max_alloc
, dmu_tx_t
*tx
)
1066 uint64_t txg
= dmu_tx_get_txg(tx
);
1067 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1069 mutex_enter(&svr
->svr_lock
);
1071 range_seg_t
*rs
= avl_first(&svr
->svr_allocd_segs
->rt_root
);
1073 mutex_exit(&svr
->svr_lock
);
1076 uint64_t offset
= rs
->rs_start
;
1077 uint64_t length
= MIN(rs
->rs_end
- rs
->rs_start
, *max_alloc
);
1079 range_tree_remove(svr
->svr_allocd_segs
, offset
, length
);
1081 if (svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] == 0) {
1082 dsl_sync_task_nowait(dmu_tx_pool(tx
), vdev_mapping_sync
,
1083 svr
, 0, ZFS_SPACE_CHECK_NONE
, tx
);
1086 svr
->svr_max_offset_to_sync
[txg
& TXG_MASK
] = offset
+ length
;
1089 * Note: this is the amount of *allocated* space
1090 * that we are taking care of each txg.
1092 svr
->svr_bytes_done
[txg
& TXG_MASK
] += length
;
1094 mutex_exit(&svr
->svr_lock
);
1096 zio_alloc_list_t zal
;
1097 metaslab_trace_init(&zal
);
1098 uint64_t thismax
= *max_alloc
;
1099 while (length
> 0) {
1100 uint64_t mylen
= MIN(length
, thismax
);
1102 int error
= spa_vdev_copy_segment(vd
,
1103 offset
, mylen
, txg
, vca
, &zal
);
1105 if (error
== ENOSPC
) {
1107 * Cut our segment in half, and don't try this
1108 * segment size again this txg. Note that the
1109 * allocation size must be aligned to the highest
1110 * ashift in the pool, so that the allocation will
1111 * not be padded out to a multiple of the ashift,
1112 * which could cause us to think that this mapping
1113 * is larger than we intended.
1115 ASSERT3U(spa
->spa_max_ashift
, >=, SPA_MINBLOCKSHIFT
);
1116 ASSERT3U(spa
->spa_max_ashift
, ==, spa
->spa_min_ashift
);
1117 thismax
= P2ROUNDUP(mylen
/ 2,
1118 1 << spa
->spa_max_ashift
);
1119 ASSERT3U(thismax
, <, mylen
);
1121 * The minimum-size allocation can not fail.
1123 ASSERT3U(mylen
, >, 1 << spa
->spa_max_ashift
);
1124 *max_alloc
= mylen
- (1 << spa
->spa_max_ashift
);
1131 * We've performed an allocation, so reset the
1134 metaslab_trace_fini(&zal
);
1135 metaslab_trace_init(&zal
);
1138 metaslab_trace_fini(&zal
);
1142 * The removal thread operates in open context. It iterates over all
1143 * allocated space in the vdev, by loading each metaslab's spacemap.
1144 * For each contiguous segment of allocated space (capping the segment
1145 * size at SPA_MAXBLOCKSIZE), we:
1146 * - Allocate space for it on another vdev.
1147 * - Create a new mapping from the old location to the new location
1148 * (as a record in svr_new_segments).
1149 * - Initiate a physical read zio to get the data off the removing disk.
1150 * - In the read zio's done callback, initiate a physical write zio to
1151 * write it to the new vdev.
1152 * Note that all of this will take effect when a particular TXG syncs.
1153 * The sync thread ensures that all the phys reads and writes for the syncing
1154 * TXG have completed (see spa_txg_zio) and writes the new mappings to disk
1155 * (see vdev_mapping_sync()).
1158 spa_vdev_remove_thread(void *arg
)
1161 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1162 vdev_copy_arg_t vca
;
1163 uint64_t max_alloc
= zfs_remove_max_segment
;
1164 uint64_t last_txg
= 0;
1166 spa_config_enter(spa
, SCL_CONFIG
, FTAG
, RW_READER
);
1167 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1168 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
1169 uint64_t start_offset
= vdev_indirect_mapping_max_offset(vim
);
1171 ASSERT3P(vd
->vdev_ops
, !=, &vdev_indirect_ops
);
1172 ASSERT(vdev_is_concrete(vd
));
1173 ASSERT(vd
->vdev_removing
);
1174 ASSERT(vd
->vdev_indirect_config
.vic_mapping_object
!= 0);
1175 ASSERT(vim
!= NULL
);
1177 mutex_init(&vca
.vca_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1178 cv_init(&vca
.vca_cv
, NULL
, CV_DEFAULT
, NULL
);
1179 vca
.vca_outstanding_bytes
= 0;
1181 mutex_enter(&svr
->svr_lock
);
1184 * Start from vim_max_offset so we pick up where we left off
1185 * if we are restarting the removal after opening the pool.
1188 for (msi
= start_offset
>> vd
->vdev_ms_shift
;
1189 msi
< vd
->vdev_ms_count
&& !svr
->svr_thread_exit
; msi
++) {
1190 metaslab_t
*msp
= vd
->vdev_ms
[msi
];
1191 ASSERT3U(msi
, <=, vd
->vdev_ms_count
);
1193 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1195 mutex_enter(&msp
->ms_sync_lock
);
1196 mutex_enter(&msp
->ms_lock
);
1199 * Assert nothing in flight -- ms_*tree is empty.
1201 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1202 ASSERT0(range_tree_space(msp
->ms_alloctree
[i
]));
1206 * If the metaslab has ever been allocated from (ms_sm!=NULL),
1207 * read the allocated segments from the space map object
1208 * into svr_allocd_segs. Since we do this while holding
1209 * svr_lock and ms_sync_lock, concurrent frees (which
1210 * would have modified the space map) will wait for us
1211 * to finish loading the spacemap, and then take the
1212 * appropriate action (see free_from_removing_vdev()).
1214 if (msp
->ms_sm
!= NULL
) {
1215 space_map_t
*sm
= NULL
;
1218 * We have to open a new space map here, because
1219 * ms_sm's sm_length and sm_alloc may not reflect
1220 * what's in the object contents, if we are in between
1221 * metaslab_sync() and metaslab_sync_done().
1223 VERIFY0(space_map_open(&sm
,
1224 spa
->spa_dsl_pool
->dp_meta_objset
,
1225 msp
->ms_sm
->sm_object
, msp
->ms_sm
->sm_start
,
1226 msp
->ms_sm
->sm_size
, msp
->ms_sm
->sm_shift
));
1227 space_map_update(sm
);
1228 VERIFY0(space_map_load(sm
, svr
->svr_allocd_segs
,
1230 space_map_close(sm
);
1232 range_tree_walk(msp
->ms_freeingtree
,
1233 range_tree_remove
, svr
->svr_allocd_segs
);
1236 * When we are resuming from a paused removal (i.e.
1237 * when importing a pool with a removal in progress),
1238 * discard any state that we have already processed.
1240 range_tree_clear(svr
->svr_allocd_segs
, 0, start_offset
);
1242 mutex_exit(&msp
->ms_lock
);
1243 mutex_exit(&msp
->ms_sync_lock
);
1246 zfs_dbgmsg("copying %llu segments for metaslab %llu",
1247 avl_numnodes(&svr
->svr_allocd_segs
->rt_root
),
1250 while (!svr
->svr_thread_exit
&&
1251 range_tree_space(svr
->svr_allocd_segs
) != 0) {
1253 mutex_exit(&svr
->svr_lock
);
1256 * We need to periodically drop the config lock so that
1257 * writers can get in. Additionally, we can't wait
1258 * for a txg to sync while holding a config lock
1259 * (since a waiting writer could cause a 3-way deadlock
1260 * with the sync thread, which also gets a config
1261 * lock for reader). So we can't hold the config lock
1262 * while calling dmu_tx_assign().
1264 spa_config_exit(spa
, SCL_CONFIG
, FTAG
);
1266 mutex_enter(&vca
.vca_lock
);
1267 while (vca
.vca_outstanding_bytes
>
1268 zfs_remove_max_copy_bytes
) {
1269 cv_wait(&vca
.vca_cv
, &vca
.vca_lock
);
1271 mutex_exit(&vca
.vca_lock
);
1274 dmu_tx_create_dd(spa_get_dsl(spa
)->dp_mos_dir
);
1275 dmu_tx_hold_space(tx
, SPA_MAXBLOCKSIZE
);
1276 VERIFY0(dmu_tx_assign(tx
, TXG_WAIT
));
1277 uint64_t txg
= dmu_tx_get_txg(tx
);
1280 * Reacquire the vdev_config lock. The vdev_t
1281 * that we're removing may have changed, e.g. due
1282 * to a vdev_attach or vdev_detach.
1284 spa_config_enter(spa
, SCL_CONFIG
, FTAG
, RW_READER
);
1285 vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1287 if (txg
!= last_txg
)
1288 max_alloc
= zfs_remove_max_segment
;
1291 spa_vdev_copy_impl(vd
, svr
, &vca
, &max_alloc
, tx
);
1294 mutex_enter(&svr
->svr_lock
);
1298 mutex_exit(&svr
->svr_lock
);
1300 spa_config_exit(spa
, SCL_CONFIG
, FTAG
);
1303 * Wait for all copies to finish before cleaning up the vca.
1305 txg_wait_synced(spa
->spa_dsl_pool
, 0);
1306 ASSERT0(vca
.vca_outstanding_bytes
);
1308 mutex_destroy(&vca
.vca_lock
);
1309 cv_destroy(&vca
.vca_cv
);
1311 if (svr
->svr_thread_exit
) {
1312 mutex_enter(&svr
->svr_lock
);
1313 range_tree_vacate(svr
->svr_allocd_segs
, NULL
, NULL
);
1314 svr
->svr_thread
= NULL
;
1315 cv_broadcast(&svr
->svr_cv
);
1316 mutex_exit(&svr
->svr_lock
);
1318 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1319 vdev_remove_complete(spa
);
1324 spa_vdev_remove_suspend(spa_t
*spa
)
1326 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1331 mutex_enter(&svr
->svr_lock
);
1332 svr
->svr_thread_exit
= B_TRUE
;
1333 while (svr
->svr_thread
!= NULL
)
1334 cv_wait(&svr
->svr_cv
, &svr
->svr_lock
);
1335 svr
->svr_thread_exit
= B_FALSE
;
1336 mutex_exit(&svr
->svr_lock
);
1341 spa_vdev_remove_cancel_check(void *arg
, dmu_tx_t
*tx
)
1343 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1345 if (spa
->spa_vdev_removal
== NULL
)
1346 return (ENOTACTIVE
);
1351 * Cancel a removal by freeing all entries from the partial mapping
1352 * and marking the vdev as no longer being removing.
1356 spa_vdev_remove_cancel_sync(void *arg
, dmu_tx_t
*tx
)
1358 spa_t
*spa
= dmu_tx_pool(tx
)->dp_spa
;
1359 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1360 vdev_t
*vd
= vdev_lookup_top(spa
, svr
->svr_vdev_id
);
1361 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
1362 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
1363 objset_t
*mos
= spa
->spa_meta_objset
;
1365 ASSERT3P(svr
->svr_thread
, ==, NULL
);
1367 spa_feature_decr(spa
, SPA_FEATURE_DEVICE_REMOVAL
, tx
);
1368 if (vdev_obsolete_counts_are_precise(vd
)) {
1369 spa_feature_decr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
1370 VERIFY0(zap_remove(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
1371 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE
, tx
));
1374 if (vdev_obsolete_sm_object(vd
) != 0) {
1375 ASSERT(vd
->vdev_obsolete_sm
!= NULL
);
1376 ASSERT3U(vdev_obsolete_sm_object(vd
), ==,
1377 space_map_object(vd
->vdev_obsolete_sm
));
1379 space_map_free(vd
->vdev_obsolete_sm
, tx
);
1380 VERIFY0(zap_remove(spa
->spa_meta_objset
, vd
->vdev_top_zap
,
1381 VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM
, tx
));
1382 space_map_close(vd
->vdev_obsolete_sm
);
1383 vd
->vdev_obsolete_sm
= NULL
;
1384 spa_feature_decr(spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
1386 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1387 ASSERT(list_is_empty(&svr
->svr_new_segments
[i
]));
1388 ASSERT3U(svr
->svr_max_offset_to_sync
[i
], <=,
1389 vdev_indirect_mapping_max_offset(vim
));
1392 for (uint64_t msi
= 0; msi
< vd
->vdev_ms_count
; msi
++) {
1393 metaslab_t
*msp
= vd
->vdev_ms
[msi
];
1395 if (msp
->ms_start
>= vdev_indirect_mapping_max_offset(vim
))
1398 ASSERT0(range_tree_space(svr
->svr_allocd_segs
));
1400 mutex_enter(&msp
->ms_lock
);
1403 * Assert nothing in flight -- ms_*tree is empty.
1405 for (int i
= 0; i
< TXG_SIZE
; i
++)
1406 ASSERT0(range_tree_space(msp
->ms_alloctree
[i
]));
1407 for (int i
= 0; i
< TXG_DEFER_SIZE
; i
++)
1408 ASSERT0(range_tree_space(msp
->ms_defertree
[i
]));
1409 ASSERT0(range_tree_space(msp
->ms_freedtree
));
1411 if (msp
->ms_sm
!= NULL
) {
1413 * Assert that the in-core spacemap has the same
1414 * length as the on-disk one, so we can use the
1415 * existing in-core spacemap to load it from disk.
1417 ASSERT3U(msp
->ms_sm
->sm_alloc
, ==,
1418 msp
->ms_sm
->sm_phys
->smp_alloc
);
1419 ASSERT3U(msp
->ms_sm
->sm_length
, ==,
1420 msp
->ms_sm
->sm_phys
->smp_objsize
);
1422 mutex_enter(&svr
->svr_lock
);
1423 VERIFY0(space_map_load(msp
->ms_sm
,
1424 svr
->svr_allocd_segs
, SM_ALLOC
));
1425 range_tree_walk(msp
->ms_freeingtree
,
1426 range_tree_remove
, svr
->svr_allocd_segs
);
1429 * Clear everything past what has been synced,
1430 * because we have not allocated mappings for it yet.
1432 uint64_t syncd
= vdev_indirect_mapping_max_offset(vim
);
1433 uint64_t sm_end
= msp
->ms_sm
->sm_start
+
1434 msp
->ms_sm
->sm_size
;
1436 range_tree_clear(svr
->svr_allocd_segs
,
1437 syncd
, sm_end
- syncd
);
1439 mutex_exit(&svr
->svr_lock
);
1441 mutex_exit(&msp
->ms_lock
);
1443 mutex_enter(&svr
->svr_lock
);
1444 range_tree_vacate(svr
->svr_allocd_segs
,
1445 free_mapped_segment_cb
, vd
);
1446 mutex_exit(&svr
->svr_lock
);
1450 * Note: this must happen after we invoke free_mapped_segment_cb,
1451 * because it adds to the obsolete_segments.
1453 range_tree_vacate(vd
->vdev_obsolete_segments
, NULL
, NULL
);
1455 ASSERT3U(vic
->vic_mapping_object
, ==,
1456 vdev_indirect_mapping_object(vd
->vdev_indirect_mapping
));
1457 vdev_indirect_mapping_close(vd
->vdev_indirect_mapping
);
1458 vd
->vdev_indirect_mapping
= NULL
;
1459 vdev_indirect_mapping_free(mos
, vic
->vic_mapping_object
, tx
);
1460 vic
->vic_mapping_object
= 0;
1462 ASSERT3U(vic
->vic_births_object
, ==,
1463 vdev_indirect_births_object(vd
->vdev_indirect_births
));
1464 vdev_indirect_births_close(vd
->vdev_indirect_births
);
1465 vd
->vdev_indirect_births
= NULL
;
1466 vdev_indirect_births_free(mos
, vic
->vic_births_object
, tx
);
1467 vic
->vic_births_object
= 0;
1470 * We may have processed some frees from the removing vdev in this
1471 * txg, thus increasing svr_bytes_done; discard that here to
1472 * satisfy the assertions in spa_vdev_removal_destroy().
1473 * Note that future txg's can not have any bytes_done, because
1474 * future TXG's are only modified from open context, and we have
1475 * already shut down the copying thread.
1477 svr
->svr_bytes_done
[dmu_tx_get_txg(tx
) & TXG_MASK
] = 0;
1478 spa_finish_removal(spa
, DSS_CANCELED
, tx
);
1480 vd
->vdev_removing
= B_FALSE
;
1481 vdev_config_dirty(vd
);
1483 zfs_dbgmsg("canceled device removal for vdev %llu in %llu",
1484 vd
->vdev_id
, dmu_tx_get_txg(tx
));
1485 spa_history_log_internal(spa
, "vdev remove canceled", tx
,
1486 "%s vdev %llu %s", spa_name(spa
),
1487 vd
->vdev_id
, (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
1491 spa_vdev_remove_cancel(spa_t
*spa
)
1493 spa_vdev_remove_suspend(spa
);
1495 if (spa
->spa_vdev_removal
== NULL
)
1496 return (ENOTACTIVE
);
1498 uint64_t vdid
= spa
->spa_vdev_removal
->svr_vdev_id
;
1500 int error
= dsl_sync_task(spa
->spa_name
, spa_vdev_remove_cancel_check
,
1501 spa_vdev_remove_cancel_sync
, NULL
, 0, ZFS_SPACE_CHECK_NONE
);
1504 spa_config_enter(spa
, SCL_ALLOC
| SCL_VDEV
, FTAG
, RW_WRITER
);
1505 vdev_t
*vd
= vdev_lookup_top(spa
, vdid
);
1506 metaslab_group_activate(vd
->vdev_mg
);
1507 spa_config_exit(spa
, SCL_ALLOC
| SCL_VDEV
, FTAG
);
1514 * Called every sync pass of every txg if there's a svr.
1517 svr_sync(spa_t
*spa
, dmu_tx_t
*tx
)
1519 spa_vdev_removal_t
*svr
= spa
->spa_vdev_removal
;
1520 int txgoff
= dmu_tx_get_txg(tx
) & TXG_MASK
;
1523 * This check is necessary so that we do not dirty the
1524 * DIRECTORY_OBJECT via spa_sync_removing_state() when there
1525 * is nothing to do. Dirtying it every time would prevent us
1526 * from syncing-to-convergence.
1528 if (svr
->svr_bytes_done
[txgoff
] == 0)
1532 * Update progress accounting.
1534 spa
->spa_removing_phys
.sr_copied
+= svr
->svr_bytes_done
[txgoff
];
1535 svr
->svr_bytes_done
[txgoff
] = 0;
1537 spa_sync_removing_state(spa
, tx
);
1541 vdev_remove_make_hole_and_free(vdev_t
*vd
)
1543 uint64_t id
= vd
->vdev_id
;
1544 spa_t
*spa
= vd
->vdev_spa
;
1545 vdev_t
*rvd
= spa
->spa_root_vdev
;
1546 boolean_t last_vdev
= (id
== (rvd
->vdev_children
- 1));
1548 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
1549 ASSERT(spa_config_held(spa
, SCL_ALL
, RW_WRITER
) == SCL_ALL
);
1554 vdev_compact_children(rvd
);
1556 vd
= vdev_alloc_common(spa
, id
, 0, &vdev_hole_ops
);
1557 vdev_add_child(rvd
, vd
);
1559 vdev_config_dirty(rvd
);
1562 * Reassess the health of our root vdev.
1568 * Remove a log device. The config lock is held for the specified TXG.
1571 spa_vdev_remove_log(vdev_t
*vd
, uint64_t *txg
)
1573 metaslab_group_t
*mg
= vd
->vdev_mg
;
1574 spa_t
*spa
= vd
->vdev_spa
;
1577 ASSERT(vd
->vdev_islog
);
1578 ASSERT(vd
== vd
->vdev_top
);
1581 * Stop allocating from this vdev.
1583 metaslab_group_passivate(mg
);
1586 * Wait for the youngest allocations and frees to sync,
1587 * and then wait for the deferral of those frees to finish.
1589 spa_vdev_config_exit(spa
, NULL
,
1590 *txg
+ TXG_CONCURRENT_STATES
+ TXG_DEFER_SIZE
, 0, FTAG
);
1593 * Evacuate the device. We don't hold the config lock as writer
1594 * since we need to do I/O but we do keep the
1595 * spa_namespace_lock held. Once this completes the device
1596 * should no longer have any blocks allocated on it.
1598 if (vd
->vdev_islog
) {
1599 if (vd
->vdev_stat
.vs_alloc
!= 0)
1600 error
= spa_reset_logs(spa
);
1603 *txg
= spa_vdev_config_enter(spa
);
1606 metaslab_group_activate(mg
);
1609 ASSERT0(vd
->vdev_stat
.vs_alloc
);
1612 * The evacuation succeeded. Remove any remaining MOS metadata
1613 * associated with this vdev, and wait for these changes to sync.
1615 vd
->vdev_removing
= B_TRUE
;
1617 vdev_dirty_leaves(vd
, VDD_DTL
, *txg
);
1618 vdev_config_dirty(vd
);
1620 spa_history_log_internal(spa
, "vdev remove", NULL
,
1621 "%s vdev %llu (log) %s", spa_name(spa
), vd
->vdev_id
,
1622 (vd
->vdev_path
!= NULL
) ? vd
->vdev_path
: "-");
1624 spa_vdev_config_exit(spa
, NULL
, *txg
, 0, FTAG
);
1626 *txg
= spa_vdev_config_enter(spa
);
1628 sysevent_t
*ev
= spa_event_create(spa
, vd
, NULL
,
1629 ESC_ZFS_VDEV_REMOVE_DEV
);
1630 ASSERT(MUTEX_HELD(&spa_namespace_lock
));
1631 ASSERT(spa_config_held(spa
, SCL_ALL
, RW_WRITER
) == SCL_ALL
);
1633 /* The top ZAP should have been destroyed by vdev_remove_empty. */
1634 ASSERT0(vd
->vdev_top_zap
);
1635 /* The leaf ZAP should have been destroyed by vdev_dtl_sync. */
1636 ASSERT0(vd
->vdev_leaf_zap
);
1638 (void) vdev_label_init(vd
, 0, VDEV_LABEL_REMOVE
);
1640 if (list_link_active(&vd
->vdev_state_dirty_node
))
1641 vdev_state_clean(vd
);
1642 if (list_link_active(&vd
->vdev_config_dirty_node
))
1643 vdev_config_clean(vd
);
1646 * Clean up the vdev namespace.
1648 vdev_remove_make_hole_and_free(vd
);
1657 spa_vdev_remove_top_check(vdev_t
*vd
)
1659 spa_t
*spa
= vd
->vdev_spa
;
1661 if (vd
!= vd
->vdev_top
)
1662 return (SET_ERROR(ENOTSUP
));
1664 if (!spa_feature_is_enabled(spa
, SPA_FEATURE_DEVICE_REMOVAL
))
1665 return (SET_ERROR(ENOTSUP
));
1668 * There has to be enough free space to remove the
1669 * device and leave double the "slop" space (i.e. we
1670 * must leave at least 3% of the pool free, in addition to
1671 * the normal slop space).
1673 if (dsl_dir_space_available(spa
->spa_dsl_pool
->dp_root_dir
,
1675 vd
->vdev_stat
.vs_dspace
+ spa_get_slop_space(spa
)) {
1676 return (SET_ERROR(ENOSPC
));
1680 * There can not be a removal in progress.
1682 if (spa
->spa_removing_phys
.sr_state
== DSS_SCANNING
)
1683 return (SET_ERROR(EBUSY
));
1686 * The device must have all its data.
1688 if (!vdev_dtl_empty(vd
, DTL_MISSING
) ||
1689 !vdev_dtl_empty(vd
, DTL_OUTAGE
))
1690 return (SET_ERROR(EBUSY
));
1693 * The device must be healthy.
1695 if (!vdev_readable(vd
))
1696 return (SET_ERROR(EIO
));
1699 * All vdevs in normal class must have the same ashift.
1701 if (spa
->spa_max_ashift
!= spa
->spa_min_ashift
) {
1702 return (SET_ERROR(EINVAL
));
1706 * All vdevs in normal class must have the same ashift
1709 vdev_t
*rvd
= spa
->spa_root_vdev
;
1710 int num_indirect
= 0;
1711 for (uint64_t id
= 0; id
< rvd
->vdev_children
; id
++) {
1712 vdev_t
*cvd
= rvd
->vdev_child
[id
];
1713 if (cvd
->vdev_ashift
!= 0 && !cvd
->vdev_islog
)
1714 ASSERT3U(cvd
->vdev_ashift
, ==, spa
->spa_max_ashift
);
1715 if (cvd
->vdev_ops
== &vdev_indirect_ops
)
1717 if (!vdev_is_concrete(cvd
))
1719 if (cvd
->vdev_ops
== &vdev_raidz_ops
)
1720 return (SET_ERROR(EINVAL
));
1722 * Need the mirror to be mirror of leaf vdevs only
1724 if (cvd
->vdev_ops
== &vdev_mirror_ops
) {
1725 for (uint64_t cid
= 0;
1726 cid
< cvd
->vdev_children
; cid
++) {
1727 if (!cvd
->vdev_child
[cid
]->vdev_ops
->
1729 return (SET_ERROR(EINVAL
));
1738 * Initiate removal of a top-level vdev, reducing the total space in the pool.
1739 * The config lock is held for the specified TXG. Once initiated,
1740 * evacuation of all allocated space (copying it to other vdevs) happens
1741 * in the background (see spa_vdev_remove_thread()), and can be canceled
1742 * (see spa_vdev_remove_cancel()). If successful, the vdev will
1743 * be transformed to an indirect vdev (see spa_vdev_remove_complete()).
1746 spa_vdev_remove_top(vdev_t
*vd
, uint64_t *txg
)
1748 spa_t
*spa
= vd
->vdev_spa
;
1752 * Check for errors up-front, so that we don't waste time
1753 * passivating the metaslab group and clearing the ZIL if there
1756 error
= spa_vdev_remove_top_check(vd
);
1761 * Stop allocating from this vdev. Note that we must check
1762 * that this is not the only device in the pool before
1763 * passivating, otherwise we will not be able to make
1764 * progress because we can't allocate from any vdevs.
1765 * The above check for sufficient free space serves this
1768 metaslab_group_t
*mg
= vd
->vdev_mg
;
1769 metaslab_group_passivate(mg
);
1772 * Wait for the youngest allocations and frees to sync,
1773 * and then wait for the deferral of those frees to finish.
1775 spa_vdev_config_exit(spa
, NULL
,
1776 *txg
+ TXG_CONCURRENT_STATES
+ TXG_DEFER_SIZE
, 0, FTAG
);
1779 * We must ensure that no "stubby" log blocks are allocated
1780 * on the device to be removed. These blocks could be
1781 * written at any time, including while we are in the middle
1784 error
= spa_reset_logs(spa
);
1786 *txg
= spa_vdev_config_enter(spa
);
1789 * Things might have changed while the config lock was dropped
1790 * (e.g. space usage). Check for errors again.
1793 error
= spa_vdev_remove_top_check(vd
);
1796 metaslab_group_activate(mg
);
1800 vd
->vdev_removing
= B_TRUE
;
1802 vdev_dirty_leaves(vd
, VDD_DTL
, *txg
);
1803 vdev_config_dirty(vd
);
1804 dmu_tx_t
*tx
= dmu_tx_create_assigned(spa
->spa_dsl_pool
, *txg
);
1805 dsl_sync_task_nowait(spa
->spa_dsl_pool
,
1806 vdev_remove_initiate_sync
,
1807 (void *)(uintptr_t)vd
->vdev_id
, 0, ZFS_SPACE_CHECK_NONE
, tx
);
1814 * Remove a device from the pool.
1816 * Removing a device from the vdev namespace requires several steps
1817 * and can take a significant amount of time. As a result we use
1818 * the spa_vdev_config_[enter/exit] functions which allow us to
1819 * grab and release the spa_config_lock while still holding the namespace
1820 * lock. During each step the configuration is synced out.
1823 spa_vdev_remove(spa_t
*spa
, uint64_t guid
, boolean_t unspare
)
1826 nvlist_t
**spares
, **l2cache
, *nv
;
1828 uint_t nspares
, nl2cache
;
1830 boolean_t locked
= MUTEX_HELD(&spa_namespace_lock
);
1831 sysevent_t
*ev
= NULL
;
1833 ASSERT(spa_writeable(spa
));
1836 txg
= spa_vdev_enter(spa
);
1838 vd
= spa_lookup_by_guid(spa
, guid
, B_FALSE
);
1840 if (spa
->spa_spares
.sav_vdevs
!= NULL
&&
1841 nvlist_lookup_nvlist_array(spa
->spa_spares
.sav_config
,
1842 ZPOOL_CONFIG_SPARES
, &spares
, &nspares
) == 0 &&
1843 (nv
= spa_nvlist_lookup_by_guid(spares
, nspares
, guid
)) != NULL
) {
1845 * Only remove the hot spare if it's not currently in use
1848 if (vd
== NULL
|| unspare
) {
1850 vd
= spa_lookup_by_guid(spa
, guid
, B_TRUE
);
1851 ev
= spa_event_create(spa
, vd
, NULL
,
1852 ESC_ZFS_VDEV_REMOVE_AUX
);
1854 char *nvstr
= fnvlist_lookup_string(nv
,
1856 spa_history_log_internal(spa
, "vdev remove", NULL
,
1857 "%s vdev (%s) %s", spa_name(spa
),
1858 VDEV_TYPE_SPARE
, nvstr
);
1859 spa_vdev_remove_aux(spa
->spa_spares
.sav_config
,
1860 ZPOOL_CONFIG_SPARES
, spares
, nspares
, nv
);
1861 spa_load_spares(spa
);
1862 spa
->spa_spares
.sav_sync
= B_TRUE
;
1864 error
= SET_ERROR(EBUSY
);
1866 } else if (spa
->spa_l2cache
.sav_vdevs
!= NULL
&&
1867 nvlist_lookup_nvlist_array(spa
->spa_l2cache
.sav_config
,
1868 ZPOOL_CONFIG_L2CACHE
, &l2cache
, &nl2cache
) == 0 &&
1869 (nv
= spa_nvlist_lookup_by_guid(l2cache
, nl2cache
, guid
)) != NULL
) {
1870 char *nvstr
= fnvlist_lookup_string(nv
, ZPOOL_CONFIG_PATH
);
1871 spa_history_log_internal(spa
, "vdev remove", NULL
,
1872 "%s vdev (%s) %s", spa_name(spa
), VDEV_TYPE_L2CACHE
, nvstr
);
1874 * Cache devices can always be removed.
1876 vd
= spa_lookup_by_guid(spa
, guid
, B_TRUE
);
1877 ev
= spa_event_create(spa
, vd
, NULL
, ESC_ZFS_VDEV_REMOVE_AUX
);
1878 spa_vdev_remove_aux(spa
->spa_l2cache
.sav_config
,
1879 ZPOOL_CONFIG_L2CACHE
, l2cache
, nl2cache
, nv
);
1880 spa_load_l2cache(spa
);
1881 spa
->spa_l2cache
.sav_sync
= B_TRUE
;
1882 } else if (vd
!= NULL
&& vd
->vdev_islog
) {
1884 error
= spa_vdev_remove_log(vd
, &txg
);
1885 } else if (vd
!= NULL
) {
1887 error
= spa_vdev_remove_top(vd
, &txg
);
1890 * There is no vdev of any kind with the specified guid.
1892 error
= SET_ERROR(ENOENT
);
1896 error
= spa_vdev_exit(spa
, NULL
, txg
, error
);
1905 spa_removal_get_stats(spa_t
*spa
, pool_removal_stat_t
*prs
)
1907 prs
->prs_state
= spa
->spa_removing_phys
.sr_state
;
1909 if (prs
->prs_state
== DSS_NONE
)
1910 return (SET_ERROR(ENOENT
));
1912 prs
->prs_removing_vdev
= spa
->spa_removing_phys
.sr_removing_vdev
;
1913 prs
->prs_start_time
= spa
->spa_removing_phys
.sr_start_time
;
1914 prs
->prs_end_time
= spa
->spa_removing_phys
.sr_end_time
;
1915 prs
->prs_to_copy
= spa
->spa_removing_phys
.sr_to_copy
;
1916 prs
->prs_copied
= spa
->spa_removing_phys
.sr_copied
;
1918 if (spa
->spa_vdev_removal
!= NULL
) {
1919 for (int i
= 0; i
< TXG_SIZE
; i
++) {
1921 spa
->spa_vdev_removal
->svr_bytes_done
[i
];
1925 prs
->prs_mapping_memory
= 0;
1926 uint64_t indirect_vdev_id
=
1927 spa
->spa_removing_phys
.sr_prev_indirect_vdev
;
1928 while (indirect_vdev_id
!= -1) {
1929 vdev_t
*vd
= spa
->spa_root_vdev
->vdev_child
[indirect_vdev_id
];
1930 vdev_indirect_config_t
*vic
= &vd
->vdev_indirect_config
;
1931 vdev_indirect_mapping_t
*vim
= vd
->vdev_indirect_mapping
;
1933 ASSERT3P(vd
->vdev_ops
, ==, &vdev_indirect_ops
);
1934 prs
->prs_mapping_memory
+= vdev_indirect_mapping_size(vim
);
1935 indirect_vdev_id
= vic
->vic_prev_indirect_vdev
;
1941 #if defined(_KERNEL) && defined(HAVE_SPL)
1942 module_param(zfs_remove_max_segment
, int, 0644);
1943 MODULE_PARM_DESC(zfs_remove_max_segment
,
1944 "Largest contiguous segment to allocate when removing device");
1946 EXPORT_SYMBOL(free_from_removing_vdev
);
1947 EXPORT_SYMBOL(spa_removal_get_stats
);
1948 EXPORT_SYMBOL(spa_remove_init
);
1949 EXPORT_SYMBOL(spa_restart_removal
);
1950 EXPORT_SYMBOL(spa_vdev_removal_destroy
);
1951 EXPORT_SYMBOL(spa_vdev_remove
);
1952 EXPORT_SYMBOL(spa_vdev_remove_cancel
);
1953 EXPORT_SYMBOL(spa_vdev_remove_suspend
);
1954 EXPORT_SYMBOL(svr_sync
);